Stateless beans in Java do not keep their state between two calls from the client. So in a nutshell we might consider them as objects with business methods. Each method takes parameters and return results. When the method is invoked some local variables are being created in execution stack. When the method returns the locals are removed from the stack and if some temporary objects were allocated they are garbage collected anyway.
From my perspective that doesn’t differ from calling method of the same single instance by separate threads. So why cannot a container use one instance of a bean instead of pooling a number of them?
Pooling does several things.
One, by having one bean per instance, you're guaranteed to be threads safe (Servlets, for example, are not thread safe).
Two, you reduce any potential startup time that a bean might have. While Session Beans are "stateless", they only need to be stateless with regards to the client. For example, in EJB, you can inject several server resources in to a Session Bean. That state is private to the bean, but there's no reason you can't keep it from invocation to invocation. So, by pooling beans you reduce these lookups to only happening when the bean is created.
Three, you can use bean pool as a means to throttle traffic. If you only have 10 Beans in a pool, you're only going to get at most 10 requests working simultaneously, the rest will be queued up.